Nonparametric adaptive age replacement

with a one-cycle criterion

P. Coolen-Schrijner,

F.P.A. Coolen

Department of Mathematical Sciences

University of Durham, Durham, DH1 3LE, UK

e-mail: Pauline.Schrijner@durham.ac.uk

Abstract
Age replacement of technical units has received much attention in the reliability
literature over the last four decades. Mostly, the failure time distribution for the units
is assumed to be known, and minimal costs per unit of time is used as optimality criterion, where renewal reward theory simplifies the mathematics involved but requires the
assumption that the same process and replacement strategy continues over a very large
(infinite) period of time. Recently, there has been increasing attention to adaptive
strategies for age replacement, taking into account the information from the process.
Although renewal reward theory can still be used to provide an intuitively and mathematically attractive optimality criterion, it is more logical to use minimal costs per
unit of time over a single cycle as optimality criterion for adaptive age replacement.
In this paper, we first show that in the classical age replacement setting, with known
failure time distribution with increasing hazard rate, the one-cycle criterion leads to
earlier replacement than the renewal reward criterion. Thereafter, we present adaptive
age replacement with a one-cycle criterion within the nonparametric predictive inferential framework. We study the performance of this approach via simulations, which
are also used for comparisons with the use of the renewal reward criterion within the
same statistical framework.

Key words: Age replacement; nonparametric predictive inference; one-cycle optimality

criterion; renewal reward theorem.
1

Introduction

Age replacement strategies for technical units are aimed at replacing such units preventively
before failure occurs, trying to avoid the need for corrective replacement which tends to be
far more expensive. This requires determination of an optimal preventive replacement time
T > 0, such that the unit will be replaced preventively if it reaches age T , or correctively
if failure occurs before it reaches age T . In the classical stochastic literature [1, 2], age
replacement is presented within a traditional Operational Research framework, with the
failure time distribution of the unit assumed to be known. If one aims at optimisation in
the sense of minimal costs per unit of time, over a very long period of time during which
the process considered does not change, and the units used all have the same failure time
distribution, then renewal reward theory [1] provides a mathematically convenient way of
computing the expected costs per unit of time. In practice, even though one acknowledges
the fact that assuming such a long period for the same process may not be realistic, the
renewal reward optimality criterion is still often considered attractive and reasonable, with
benefit of its mathematical simplicity.
During the last decade [3, 4, 5], researchers have become interested in so-called adaptive
replacement strategies, where instead of assuming complete knowledge of a units failure
time distribution, one uses information from the process to learn about this distribution.
For example, the Bayesian statistical framework [3, 4, 5] allows the use of a parametric
distribution, with an assumed prior distribution reflecting subjective information available
before the process starts, and which, combined with process information, gives a posterior
distribution, which can be used to determine the optimal replacement strategy for the next
unit in the process. As discussed by Mazzuchi and Soyer [3], such an adaptive approach
undermines the use of the renewal reward criterion as one explicitly does not wish to use the
same replacement strategy over a very long period of time. Hence, they proposed the use of
a one-cycle optimality criterion, aiming at minimal expected costs per unit of time for the
cycle corresponding to the unit considered.
Throughout this paper, the term cycle is used for the period starting at the moment
that a new unit starts functioning, and ending at the moment that the next unit has been
installed, following either preventive or corrective replacement. The time required for any replacement is assumed to be neglectible throughout (adapting to include a known replacement
time is relatively straightforward), and the costs of preventive replacement and corrective
replacement are assumed to be known constants, where the costs per cycle include the costs

incurred by the replacement at the end of a cycle.

Recently, we have introduced and studied the use of nonparametric predictive inference
(NPI) for adaptive age replacement strategies [6, 7], leading to optimal strategies which
are fully adaptive to process data without the need to assume a parametric probability
distribution for a units failure time. So far, we studied NPI-based age replacement using
the renewal reward criterion, in this paper we follow Mazzuchi and Soyers approach by
adopting the one-cycle optimality criterion.
For a detailed introduction to NPI in reliability we refer to Coolen, et al. [8], here we
briefly give the main concept. Consider the situation where we have observed n failure times
for units in the same process, ordered as 0 < x(1) < x(2) < . . . < x(n) (we assume throughout
this paper that there are no ties in the observed failure times). Instead of assuming a
known probability distribution for the failure time Xn+1 of the next unit, we directly specify
probabilities for Xn+1 according to Hills assumption A(n) [9, 10], that is,
P (Xn+1 (x(j) , x(j+1) )) =

1
, j = 0, . . . , n,
n+1

(1)

where x(0) = 0 and x(n+1) = r, where r is either infinity or a finite upper bound for the
support of Xn+1 . For a further discussion of A(n) , and historical references which provide
justifications of this statistical approach, see [8, 10]. As a consequence of (1), the predictive
survival function for Xn+1 , at the times x(j) , is equal to
SXn+1 (x(j) ) =

n+1j
, j = 0, . . . , n + 1.
n+1

(2)

Without further assumptions it is not possible to give a precise value for this survival function
at other times, as A(n) assigns probability mass 1/(n + 1) to the open intervals created by
the observed failure times, but does not put any further restriction on the distribution of the
probability mass within these intervals. Straightforward bounds for this survival function
can be derived at all times [8], these are upper and lower survival functions within the theory
of interval probability [10]. These upper and lower survival functions can themselves be used,
without further assumptions, as was e.g. done for our NPI-based age replacement strategies
using the renewal reward criterion [6, 7]. In this paper, we will instead introduce a further
assumption on the distribution of the probability mass within these intervals, creating a
precise survival function for Xn+1 , as we will explain in Section 3.
In Section 2, we compare the optimal age replacement strategies corresponding to the
renewal reward criterion and the one-cycle criterion from the classical stochastic perspective,
with the failure time distribution assumed to be known. We will show that, for distributions
3

with increasing hazard rates, optimal preventive replacement based on the one-cycle criterion
will take place earlier than when based on the renewal reward criterion. As our NPI-based
method does not assume increasing hazard rates, this result does not apply automatically for
NPI-based age replacement with a one-cycle criterion, as presented in Section 3. In Section 4
we illustrate our method via a small example, including comparison with our NPI-based age
replacement using the renewal reward criterion [6, 7]. In Section 5 we present and discuss
the results from a simulation study, which provide further insights in our adaptive method.
We end the paper with some concluding remarks in Section 6, and an appendix containing
the proofs of two lemmas presented in Section 3.

Classical age replacement

In this section, we consider the classical stochastic model for age replacement [1, 2], where
the failure time of the unit is assumed to be an absolutely continuously distributed random
quantity X 0 with known probability distribution, with cumulative distribution function
F (x) = P (X x), probability density function (pdf) f (x), hazard rate h(x) = f (x)/(1
F (x)), and expected value E(X). To avoid mathematical complications, we assume that
F (0) = 0, F (x) > 0 for all x > 0, and E(X) < . The costs included in the age replacement
model are assumed to be known constants, with cost of preventive replacement c1 and cost
of corrective replacement c2 , where c2 > c1 > 0.
The renewal reward criterion minimises the expected costs per unit of time, where the
same replacement strategy is assumed to be used over an infinite period of time, applied to a
sequence of units whose failure time random quantities all have the same known probability
distribution. Application of the renewal reward theorem [1] implies that the cost function,
for replacement strategy T > 0 (i.e. replacement at time T , or at failure if that occurs before
T ), equals the expected costs per cycle divided by the expected length of a cycle,
Cr (T ) =

c1 (1 F (T )) + c2 F (T )
.
RT
(1 F (x))dx
0

(3)

Let the optimal strategy be denoted by Tr (which might be equal to infinity), then the
first-order necessary condition for optimality (e.g. [11]) of a finite Tr is that it must satisfy
Z T
c1
.
(4)
h(T )
(1 F (x))dx F (T ) =
c2 c 1
0

We define

gr (T ) = h(T )

(1 F (x))dx F (T ).

(5)

Clearly, gr (T ) is a continuous function with gr (0) = 0, and, if h() = limT h(T ) exists,
then
lim gr (T ) = h()E(X) 1.

(6)

If h(T ) is monotonously strictly increasing, then gr (T ) is also strictly increasing as

dgr (T )
= h0 (T )
dT

(1 F (x))dx

(7)

is positive if h0 (T ) > 0 for all T > 0. Hence, for monotonously strictly increasing h(T ),
the condition gr (T ) =

c1
c2 c1

either has no solution, if h()

c2
,
(c2 c1 )E(X)

in which case one

would not replace the unit preventively at a finite time, or it has a unique finite solution Tr ,
if h() >

c2
.
(c2 c1 )E(X)

If there is no finite optimal preventive replacement time, then the cost function obtains
its minimal value in the limit,
Cr () = lim Cr (T ) =
T

c2
.
E(X)

(8)

Hence, if there is a finite optimal preventive replacement strategy, the minimal expected costs
per unit of time, over an infinite time horizon and based on the renewal reward criterion, is
less than

c2
.
E(X)

The one-cycle criterion [3] minimises the expected costs per unit of time for a single
cycle. This criterion has been mostly neglected in the literature, which may be caused by the
fact that, traditionally, one assumes a known probability distribution for the units failure
time, in which case it is logical to use the same optimal replacement strategy over many
cycles. However, when aiming at adaptive age replacement strategies, where the optimal
replacement time can change per cycle, the one-cycle criterion is particularly attractive, as
such changing replacement times are explicitly excluded by the underlying justification for
use of the renewal reward criterion. Mazzuchi and Soyer [3] proposed this one-cycle criterion
for Bayesian adaptive age replacement strategies. The cost function, as function of the
random failure time X, for replacement strategy T , is the costs per unit of time during one
cycle,
C1 (X, T ) =

c2 /X

if X < T,

c1 /T

if X T.

(9)

To avoid mathematical complexity for one-cycle optimality, we assume that E(1/X) exists,
which is a condition on the failure time distribution for X for values close to 0 (which
particularly excludes the use of the Exponential distribution close to 0 in what follows). The
5

one-cycle optimality criterion is minimisation of the expected value of C1 (X, T ) with regard
to the probability distribution for X,
C1 (T ) = E(C1 (X, T )) = c2

T
0

c1
1
f (x)dx + (1 F (T )).
x
T

(10)

The optimal strategy T1 , corresponding to this criterion, might be infinite, but if a finite
optimum exists it must satisfy the first-order necessary condition for optimality, which leads
to
T h(T ) =

c1
.
c2 c 1

(11)

We define
g1 (T ) = T h(T ),

(12)

which is clearly a continuous function with g1 (0) = 0, and which is monotonously strictly
increasing with no upper bound for T if h(x) is monotonously strictly increasing.
Hence, for failure time distributions with such hazard rates, a unique finite optimal preventive
replacement time T1 exists according to this one-cycle criterion, and this is derived via
g1 (T1 ) =

c1
.
c2 c1

If the hazard rate h(x) is such that g1 (T ) <

c1
c2 c1

for all finite T , which

is unlikely to be the case in relevant replacement situations as it relates to a very strong

wear-in effect, then the cost function would have minimal value in the limit,
C1 () = lim C1 (T ) = c2 E(1/X).
T

(13)

Hence, in most situations of interest for possible preventive replacement, the minimal expected costs per unit of time when considered over one cycle is less than c2 E(1/X).
We can now compare the optimal age replacement strategies Tr and T1 , corresponding
to the renewal reward criterion and the one-cycle criterion, respectively. If the probability
distribution of a units absolutely continuously distributed failure time X > 0 is such that
E(X) and E(1/X) are both finite, and that it has a monotonously strictly increasing hazard
rate h(x), then
T1 < T r .

(14)

This follows immediately from the fact that, for all T > 0,
gr (T ) < g1 (T ),
which easily follows from

RT
0

(15)

(1 F (x))dx < T and F (T ) > 0, and the fact that these

functions are continuous and gr (0) = g1 (0) = 0. Of course, Tr can be infinite as discussed
6

above. Hence, in the classical stochastic setting with known probability distribution for a
units failure time, if the unit is subject to wearout in the sense of a strictly increasing hazard
rate for its failure time, then the optimal age replacement strategy according to the one-cycle
criterion leads to earlier preventive replacement than the optimal strategy according to the
renewal reward criterion.

NPI age replacement with a one-cycle criterion

In this section, we present NPI-based adaptive age replacement for unit n + 1, using the
one-cycle criterion, hence aiming at minimisation of the cost function C1 (T ), as given by
(10), for random failure time Xn+1 . This provides an intuitively attractive alternative to
NPI-based age replacement using the renewal reward criterion, as presented in [6, 7].
In [6, 7], we did not add any further assumptions on the probability distribution for Xn+1
to the inferences which follow directly from A(n) , using both the upper and lower survival
functions for Xn+1 , leading to two optimal age replacement strategies which could be different. The emphasis there was not on such differences, but on how well these strategies adapt
to underlying distributions, which was studied via simulations and turned out to be pretty
impressive already for sample sizes of n = 10, when sampling from Weibull distributions
with increasing hazard rate.
If we attempt to follow the same approach in this paper, we run into difficulty when using
the lower survival function [8] for Xn+1 , as this puts probability 1/(n + 1) adherent to 0,
which implies that the expected value of 1/Xn+1 is not finite. To overcome this problem, we
use a pragmatic solution by assuming, in addition to A(n) , a particular manner in which the
probability masses 1/(n + 1) are divided over the intervals created by the n observed failure
times. Of course, this additional assumption will influence the resulting strategies, but this
influence will decline with increasing n. An advantage of this pragmatic approach is that we
only need to study our approach for a single survival function related to n observed failure
times, while keeping the A(n) assumption.
In this paper, we assume, for NPI-based age replacement with a one-cycle criterion, in
addition to A(n) , that the probability masses 1/(n+1) assigned to the open intervals between
consecutive observed failure times x(1) , . . . , x(n) are uniformly distributed per interval. The
two end intervals of this partition of the real axis require special attention. The probability
mass in the interval beyond x(n) can only be distributed uniformly by assuming a finite
end-point r for the support of Xn+1 . Alternatively, we could have assumed a particular
7

distribution of the probability mass 1/(n + 1) over (x(n) , ). However, from the expected
costs C1 (T ), as presented in Section 2, it is clear that the precise distribution of probability
mass beyond preventive replacement age T is irrelevant. In situations where the data indicate
that preventive replacement may actually be useful, one expects that the optimal preventive
replacement age T is less than the largest observed failure time x(n) . Therefore, in this paper
we assume a finite r and distribute the probability mass 1/(n + 1) uniformly over (x(n) , r),
where of course we always take r larger than x(n) . If the optimal replacement time over (0, r]
would occur beyond x(n) , the data would clearly not support preventive replacement.
For the interval (0, x(1) ), a Uniform distribution of the probability mass 1/(n + 1) cannot
be assumed together with the one-cycle criterion, as the integral of f (x)/x would still not
converge over a small interval immediately to the right of 0. The same problem would also
remain with an Exponential distribution over this interval. Hence, we assume that this
probability mass 1/(n + 1) is distributed over (0, x(1) ) according to a Gamma distribution
with shape parameter 2, and scale parameter chosen such that
Zx(1)
2 tet dt =

1
,
n+1

(16)

so is calculated via the equation

n
= ex(1) (1 + x(1) ).
n+1

(17)

The assumption A(n) , and these additional assumptions for the distribution of the probability
masses 1/(n+1) over the intervals created by the observed failure times, lead to the following
specification of the predictive survival function for the failure time Xn+1 of unit n + 1 within
these intervals, to be used together with the A(n) -based values at the observed failure times
as given by (2) (with x(n+1) = r as before),
SXn+1 (x) = 1

Zx

tet dt, x (0, x(1) ),

(18)


x x(j)
SXn+1 (x) =
SXn+1 (x(j) ) +
SXn+1 (x(j+1) )
x(j+1) x(j)


x(j+1) x
1
= SXn+1 (x(j+1) ) +
, x (x(j) , x(j+1) ), j = 1, . . . , n. (19)
n + 1 x(j+1) x(j)


x(j+1) x
x(j+1) x(j)

Using this predictive survival function for Xn+1 , and the fact that the pdf fXn+1 for Xn+1
satisfies fXn+1 (x) = dSXn+1 (x)/dx, we obtain the following expression for the pdf for Xn+1
8

which we use in our NPI-based age replacement with a one-cycle criterion in this paper,



1
1

, x (x(j) , x(j+1) ), j = 1, . . . , n,
n + 1 x(j+1) x(j)
fXn+1 (x) =
(20)

2 x
xe , x (0, x(1) ).
With this pdf, the expected costs per unit time for NPI-based age replacement of unit

n + 1, using a one-cycle criterion, and based on ordered observed failure times x(1) , . . . , x(n) ,
is

ZT

Cone (T ) = EXn+1 [C1 (Xn+1 , T )] =

c2
fX (x)dx +
x n+1

c1
fX (x)dx.
T n+1

(21)

Throughout this paper, we use the notation Cone (T ) for the one-cycle criterion cost function
in case of our NPI-based age replacement, with the additional assumptions leading to pdf
fXn+1 as in (20). Substituting this pdf into (21) yields Lemma 1, the proof is given in the
Appendix.

Lemma 1
a. For T = x(j) and j = 1, . . . , n + 1, we have
1
Cone (x(j) ) = c2 (1 ex(1) ) +
n+1

c2

j1
X
ln(x(l+1) ) ln(x(l) )
l=1

x(l+1) x(l)

+ c1

n+1j
x(j)

)

(22)

b. For T (x(j) , x(j+1) ) and j = 1, . . . , n, we have

1
Cone (T ) = c2 (1 ex(1) ) +
n+1

c2

j1
X
ln(x(l+1) ) ln(x(l) )
l=1

!
ln(T ) ln(x(j) )
+
x(l+1) x(l)
x(j+1) x(j)


x(j+1) T
c1
+
+nj
.
T x(j+1) x(j)
(23)

c. For T (0, x(1) ), we have

Cone (T ) = c2 (1 e



1
1
c1 n
c1
T
x(1)
(x(1) + ) +
(T + ) e
. (24)
)+ e
T

T n+1

It is easy to show that the cost function Cone (T ), as given in Lemma 1, is continuous
for all T . Lemma 2, of which the proof is also given in the Appendix, uses the fact that if
the ratio c2 /c1 satisfies a certain condition, then Cone (T ) is first decreasing and thereafter
increasing over (x(j) , x(j+1) ), j = 0, . . . , n, so that Cone (T ) has a unique local minimum over
each interval (x(j) , x(j+1) ).

Lemma 2
a. For each interval (x(j) , x(j+1) ), j = 1, . . . , n, if
x(j+1) + (n j)(x(j+1) x(j) )
x(j+1) + (n j)(x(j+1) x(j) )
c2
<
<
x(j+1)
c1
x(j)
then Cone (T ) has a local minimum in
c1
j = (x(j+1) + (n j)(x(j+1) x(j) )).
c2

(25)

(26)

b. For the interval (0, x(1) ), if

0 =

c1 +

c1 (4c2 3c1 )
< x(1)
2(c2 c1 )

(27)

then Cone (T ) has a local minimum in 0 .

Denote by Tj the value of T that minimises Cone (T ) over the interval [x(j) , x(j+1) ], that
is,
Tj = argmin{Cone (T ), T [x(j) , x(j+1) ], j = 0, . . . , n}.

(28)

As a result of Lemma 2 we now obtain the following analytical expressions for Tj .

A. Let
a=

x(j+1) + (n j)(x(j+1) x(j) )

x(j+1) + (n j)(x(j+1) x(j) )
, and b =
.
x(j+1)
x(j)

(29)

For each interval [x(j) , x(j+1) ], j = 1, . . . , n, the value of T that minimises Cone (T ) for
unit n + 1 is

c2

x(j)
if
b,

c1

c2
a,
x(j+1)
if
Tj =
c1

c1
c2

< b.
(x(j+1) + (n j)(x(j+1) x(j) )) if a <
c2
c1
10

(30)

B. For the interval [0, x(1) ], Cone (T ) is minimal for

p
p

c1 + c1 (4c2 3c1 )
c1 + c1 (4c2 3c1 )

if
< x(1) ,

2(c2 c1 )
2(c
2 c1 )
p
T0 =

c1 + c1 (4c2 3c1 )

x(1)
if
x(1) .
2(c2 c1 )

(31)

As a consequence, the global minimum Tone

of the expected costs per unit time for our

NPI-based age replacement with a one-cycle criterion, over the interval [0, r] satisfies

Tone
= argmin{Cone (Tj ), j = 0, . . . , n}.

(32)

Computations involved in deriving this Tone

, for given ordered failure times x(1) , . . . , x(n) and

assumed value of r > x(n) , are straightforward, as every expression is available analytically,
except for computation of the scale parameter determining the specific Gamma distribution
used for the interval (0, x(1) ), which is easily derived numerically from (17).

Example

Suppose we have observed five failure times: 4, 6, 10, 11 and 15. Each preventive replacement
costs c1 = 1, while each corrective replacement costs c2 = 10. Assume that an upper bound
for the support of X6 is given by r = 25. We want to find the NPI-based optimal age
replacement time for unit 6 in the sense of minimising Cone (), so based on the one-cycle
criterion, and using A(5) , together with uniformly distributed probability masses within
each of the intervals (x(1) , x(2) ), . . . , (x(5) , r), Gamma distributed probability mass within the
interval (x(0) , x(1) ), and the observed failure times. Applying (30) and (31) yields the local
optima Ti , and the corresponding values of Cone (Ti ), per interval, as given in Table 1. Hence,

the optimal age replacement time is Tone

= 2.153, with corresponding Cone (Tone
) = 1.0312.

We compare this optimal strategy with the optimal age replacement time in case the
cost function is based on a renewal argument [6, 7]. Using a renewal argument, the long-run
expected costs per unit time, for unit n + 1, are Cr (T ) as given by (3), which can also be
written as
Cr (T ) =

c2 S(T )(c2 c1 )
.
RT
S(x)dx
0

11

(33)

Ti

Cone (Ti )

[0, 4]

2.153 1.0312

[4, 6]

1.1561

[6, 10]

1.3968

[10, 11] 10

1.5485

[11, 15] 11

1.6877

[15, 25] 15

1.7977

Table 1: Local optima Ti per interval, with corresponding values Cone (Ti )

In NPI-based age replacement with the renewal criterion [6, 7], we used A(n) without adding
further assumptions on the distributions of the probability masses per interval. A lower
(upper) bound for the survival function is then obtained by shifting all the probability mass
in the interval in which x lies to the left (right) end-point of the interval, see [6] for more
details. These lower and upper bounds for the survival function lead to upper and lower
bounds, respectively, for the long-run expected costs per unit time, also called the renewal
upper and lower cost functions. From [6] we know that, for this example, minimising the
renewal upper cost function yields an optimal age replacement time of 4, with corresponding
upper cost function value of 0.75, while minimising the renewal lower cost function yields an
optimal age replacement time of just before 4 (see [6] for further details), with corresponding
cost function value of 0.25. Figure 1 shows the cost function based on a one-cycle criterion,
together with the renewal upper and lower cost functions. The optimal age replacement time
based on a one-cycle criterion is less than the optimal age replacement times corresponding
to both the renewal upper and lower cost functions. This is in line with our result (14),
which was proven for the classical situation with the probability distribution for the failure
time assumed known, and with increasing hazard rate. Hence, the result (14) does not
apply in our NPI-based setting, as we do not add any assumption on the hazard rate in our
renewal reward setting. In our simulation study, presented in Section 5, we also consider
this comparison in detail. Figure 1 also illustrates that, in this example, the one-cycle cost
function lies above the renewal upper and lower cost functions for larger values of T . We
also consider this further in Section 5.

12

8
4
0

Costs per unit time

Onecycle cost function

Renewal upper cost function
Renewal lower cost function

10

15

Figure 1: Cost functions for one-cycle and renewal criteria

Simulations

In this section we present results from simulation studies, to illustrate our NPI-based method
and to discuss several of its features. The simulations and computations were performed with
the statistical package R [12]. The failure times are simulated from a known probability distribution, where we have restricted attention to the Weibull distribution with scale parameter
1 and shape parameter (denoted by W (, 1)), which has pdf

fX (t) = t1 e(t ) , t 0.

(34)

This enables us to compare our simulation results for the NPI-based optimal replacement

times according to the one-cycle criterion, denoted by Tone

, with the theoretical optimal

13

replacement times according to the same criterion but with the known Weibull distribution,
which we denote by T1 , and which are obtained by minimisation of the cost function C1 (T )
as given by (10).
As in the example in Section 4, we also compare our results under the one-cycle criterion
with the results under the renewal reward criterion, by calculating the NPI-based optimal
replacement times for the renewal lower and upper cost functions, denoted by T ren and

T ren , respectively. We denote by Tr the theoretical optimal age replacement time under the
renewal reward criterion for the known probability distribution for X, which is obtained by
minimisation of the cost function Cr (T ) as given by (3).
For as far as the replacement costs are concerned, only the ratio c2 /c1 is relevant for the
minimum of the cost function, so we have used c1 = 1, without loss of generality. We have
run simulations with c2 equal to 10 or 50. The number of initially observed failure times, n,
equals 10 or 50, and in each case we have simulated 1000 times. An upper bound r = 5 for
the support of Xn+1 is assumed, we checked to ensure that all simulated observations where
less than 5, which was indeed the case.
Studying the simulation results below, great care must be taken with notation, as we
compare two different optimality criteria, for both we consider the NPI-based cost functions
and optima as well as the theoretical ones with the Weibull distribution assumed to be
known, and we also consider the theoretical cost function values at the NPI-based optimal
replacement times, in order to compare the effectiveness of our NPI-based method with
the underlying theoretical model. Throughout, we continue to use notation as introduced
above. For our NPI-based approach, the one-cycle cost functions and corresponding optimal
replacement times have an index one, whereas the corresponding NPI-based renewal cost
functions and optimal replacement times have an index ren. The theoretical cost functions
and optima, related to the one-cycle criterion, have an index 1, those related to the renewal
criterion have an index r. This leads to the following notation.
The optimal theoretical one-cycle expected age replacement costs per unit time are
C1 (T1 ), and the optimal theoretical long-run renewal expected costs per unit time are
Cr (Tr ). Within our NPI-based approach, so explicitly in terms of unit n + 1, the opti
mal expected age replacement costs per unit time with the one-cycle criterion is Cone (Tone
).

Using similar notation, C ren (T ren ) and C ren (T ren ) are the NPI-based optimal renewal upper
and lower long-run expected costs per unit time, respectively. For comparisons in our simu

lations, we use one = (C1 (Tone

) C1 (T1 ))/C1 (T1 ), ren = (Cr (T ren ) Cr (Tr ))/Cr (Tr ) and

ren = (Cr (T ren ) Cr (Tr ))/Cr (Tr ). These 0 s indicate how good our NPI-based optimal
14

W (2, 1)

W (3, 1)

c2 = 10 c2 = 50 c2 = 10 c2 = 50
T1

0.2357

0.1010

0.3333

0.1895

C1 (T1 )

8.6416 19.8663

4.5332

7.9272

C1 () 17.7245 88.6227 13.5412 67.7059

1 ()

1.0511

3.4610

1.9871

7.5410

Tr

0.3365

0.1431

0.3825

0.2170

Cr (Tr )

6.0561 14.0239

3.9494

6.9215

Cr () 11.2838 56.4190 11.1985 55.9923

r ()

0.8632

3.0231

1.8355

7.0896

Table 2: Theoretical results

replacement times are, compared to the corresponding theoretical optima. Clearly, a low
value of such a indicates that our method performs well, in the sense that our optimal
NPI-based replacement time, which is based directly on the simulated set of failure data,
provides a replacement strategy which is nearly as cost effective as the theoretically best
strategy, which of course in practice is not known as the underlying failure time distribution will not be known. Hence, these s are useful performance measures for our method,
illustrating the joint effects of how well our NPI-based age replacement strategies adapt to
the data, and how well the data capture the characteristics of the underlying failure time
distribution for as far as relevant to these inferences.
Table 2 gives the theoretical optimal replacement times T1 and Tr , and the corresponding
minimal cost function values C1 (T1 ) and Cr (Tr ), for the failure time distributions W (2, 1)
and W (3, 1). We have also included the limiting values of these cost functions for T ,
denoted by C1 () and Cr (), which relate to no preventive replacement being carried out,
and the values 1 () = (C1 () C1 (T ))/C1 (T ) and r () = (Cr () Cr (Tr ))/Cr (Tr ),
which are indicators of the theoretical losses, relative to the optimal costs, if no preventive
replacements were carried out. The values of 1 () and r () are larger for W (3, 1)
than for W (2, 1), which indicates that preventive replacement is more effective for W (3, 1)
than for W (2, 1). This results from the fact that the variance of the W (3, 1) distribution is
smaller than the variance of W (2, 1), and replacement policies tend to be more effective if
the underlying distribution has a smaller variance [6, 7]. Tables 3 and 4 give the simulation
results for W (2, 1) and W (3, 1), respectively. Tables 5 and 6 show the positions of the
15


Tone

Cone (Tone
)

one

C ren (T ren )

T ren

T ren

ren

C ren (T ren )

ren

CASE 1-1: W (2, 1), c2 = 10 and n = 10

Mean

0.2669

16.2644

0.0989 0.5454

6.5055

0.1357 0.4793

3.7927

0.0974

Median 0.2526

10.7120

0.0438 0.4978

6.1994

0.0809 0.3949

3.4985

0.0437

Sd

48.7599

0.1698 0.2336

2.1377

0.1545 0.5118

1.5494

0.1418

0.1137

CASE 1-2: W (2, 1), c2 = 10 and n = 50

Mean

0.2602

11.0495

0.0416 0.3921

5.7470

0.0450 0.3613

5.0962

0.0438

Median 0.2552

9.6318

0.0188 0.3715

5.7014

0.0214 0.3438

5.0497

0.0213

Sd

6.1479

0.0598 0.1204

1.0545

0.0605 0.1138

1.0055

0.0579

0.0733

CASE 1-3: W (2, 1), c2 = 50 and n = 50

Mean

0.1078

34.1054

0.1334 0.2375

14.2651

0.1651 0.1836

8.6752

0.1020

Median 0.0999

24.4769

0.0605 0.2244

13.4647

0.1024 0.1737

7.9462

0.0553

Sd

37.3376

0.2226 0.0845

4.5893

0.1808 0.0742

3.7394

0.1265

0.0491

Table 3: Simulation results for W (2, 1)

individual values Tone

in relation to the corresponding values T1 , T ren and T ren .

The main conclusions that can be drawn from these simulation results are discussed
below. We also pay attention to the results based on the one-cycle criterion versus those
for the renewal criterion. We refer to [6, 7] for conclusions on such results from similar
simulation studies for the renewal criterion only.

Conclusions from simulations:

1. Tables 3 and 4 show that the means and medians of Tone

are smaller than the means and

medians of T ren and T ren , which is in line with result (14) (even though, as discussed
before, (14) does not apply here). Tables 5 and 6 show that this not only holds for the
means and medians, but for all individual simulated data sets in our study, although
in several situations the optima according to the different criteria are equal (see point

7 below). The standard deviations of Tone

are also smaller than the corresponding

ones for the renewal criterion, indicating less spread in the optima, in particular fewer
relatively large values occur when the one-cycle criterion is used.

16


Tone

Cone (Tone
)

one

C ren (T ren )

T ren

ren

T ren

C ren (T ren )

ren

CASE 2-1: W (3, 1), c2 = 10 and n = 10

Mean

0.3484

8.1485

0.1225 0.5475

4.8295

0.2064 0.4843

2.6125

0.1529

Median 0.3397

6.9477

0.0609 0.5429

4.6075

0.1343 0.4764

2.4340

0.0888

Sd

4.7086

0.1850 0.1527

1.4773

0.2271 0.1518

0.9619

0.1814

0.1160

CASE 2-2: W (3, 1), c2 = 10 and n = 50

Mean

0.3660

5.6566

0.0565 0.4338

3.8672

0.0589 0.4139

3.3477

0.0535

Median 0.3577

5.2749

0.0249 0.4297

3.8237

0.0315 0.4037

3.2960

0.0276

Sd

1.9090

0.0772 0.0919

0.6792

0.0741 0.0920

0.6261

0.0685

0.0814

CASE 2-3: W (3, 1), c2 = 50 and n = 50

Mean

0.1815

14.6619

0.1356 0.3080

8.3397

0.2185 0.2723

4.5780

0.1590

Median 0.1784

12.2366

0.0558 0.2995

8.0505

0.1185 0.2670

4.2240

0.0799

Sd

9.1017

0.2515 0.0819

2.3185

0.2563 0.0835

1.6400

0.2036

0.0613

Table 4: Simulation results for W (3, 1)

2. Tables 3 and 4, when compared to Table 2, show that the means and medians of

Tone
are pretty close to the corresponding T1 , which indicates that our method adapts

well to the underlying distribution for data sets of size 10 and 50, as used in this
simulation study. For the corresponding values based on the renewal criterion, there is
a noticeable improvement in the adaptation to the underlying distribution for n = 50
over n = 10. From this study, it seems that our NPI-based method adapts better to
the underlying distribution for the one-cycle criterion than for the renewal criterion, in
particular for smaller values of n (see also points 5 and 6 below). Tables 5 and 6 show

that, for individual cases, Tone

can be smaller or larger than the theoretical T1 based

on the underlying failure time distribution. This is caused by our NPI-based method
adapting well to the sampled data, and the randomness in such simulated data.

3. The means and medians of Cone (Tone

) are larger than C1 (T1 ). However, for individual

cases (not shown), Cone (Tone

) can also be smaller than C1 (T1 ). The means and medians

of Cone (Tone
) are also larger than the corresponding values of C ren (T ren ), C ren (T ren ),

and Cr (Tr ).

17

Case 1-1
T1

T ren

> 562
Tone

Tone
=

Tone

Case 1-2

Case 1-3

T1

T ren

T ren

T1

0 607

495

324

451

< 438 1000 1000 393

676

549 505 1000 1000

T ren

T ren T ren

Table 5: The positions of the NPI-based Tone

for W(2,1)

Case 2-1

Case 2-2

T ren

Tone
> 514

0 632

Tone
=

Tone

Case 2-3

T ren

T ren

437

562

676

< 486 1000 1000 368

438

324 563 1000 1000

T ren

T T ren T ren

Table 6: The positions of the NPI-based Tone

for W(3,1)

4. The medians of Tone

are smaller than the corresponding means, which is due to the

fact that the distributions of the Tone

s are skewed to the right. In our earlier work on

NPI-based age replacement with the renewal criterion [6, 7], this was also discussed,
and we regard the medians as better indicators of the performance of our methods
than the means.
5. The means and medians of one are pretty small, indicating that our method tends to
adapt well to the data, in the sense that the resulting NPI-based optimal replacement
times would not lead to large losses compared to the theoretically optimal replacement
times. The distributions of the values one from our simulations are highly skewed to
the right, meaning that many values are very close to 0, but some are large (which is
shown both by the large differences between corresponding means and medians, and
the large standard deviations). Due to the nature of the one-cycle cost function, large
values of one occur particularly if the NPI-based method suggests relatively small
preventive replacement times, which occurs if the simulated data happens to include
small failure times. For increasing n (compare Cases 1-1 and 1-2, and Cases 2-1 and 22), the one tend to decrease, which shows that our method provides better adaptation
to the underlying distribution for larger values of n. In particular, for larger n the effect
18

of very small simulated data values decreases, so the NPI-based preventive replacement
times do not become very small as easily as for n = 10.
6. Comparing Cases 1-2 and 1-3, and Cases 2-2 and 2-3, we see that the means and
medians of one increase with c2 , and also the standard deviations increase noticeably.
For larger corrective replacement costs, one is more careful and tends to plan earlier
preventive replacement. Therefore, if the simulated data include one or more very
small observed failure times, the NPI-based age replacement time is also very small,
leading to large one .

7. If n is small or c2 /c1 is large (Cases 1-1, 1-3, 2-1 and 2-3), Tone
is smaller than T ren

and T ren (see Tables 5 and 6). However, if n is not too small and c2 /c1 not too large,

Tone
is quite often equal to T ren or T ren (Cases 1-2 and 2-2). This follows from (30)

and (31) and the fact that the renewal upper and lower cost functions are minimised
in the right end-point of each interval (or just before it, see [6]).

Concluding Remarks

The method presented in this paper is fully adaptive to available failure data, but therefore
requires availability of such data, which may often be unrealistic. However, one should
not forget that the classical method assumes full knowledge of an underlying failure time
distribution, which is even more far stretched. Of course, if such failure data are available, our
method is of direct use, either on its own or, which may be more attractive, in combination
with more established methods for age replacement [1] or alternative adaptive methods
[3, 4, 5]. If different methods result in pretty similar preventive replacement times, the
data strongly support the additional assumptions underlying the other methods, and it is
straightforward to decide on a good replacement policy. If, however, these different methods
lead to rather different optimal replacement times, our method has the advantage of being
based on quite minimal modelling assumptions, so the resulting strategies from the other
models are more influenced by modelling assumptions, which one should then try to verify
or study in more detail. Our method can also be generalized, relatively straightforwardly, to
situations where the observed data include some right-censored observations, which may for
example result from earlier preventive replacements. This can be achieved by replacing A (n)
by the generalized assumption presented by Coolen and Yan [13], which similarly provides
predictive probabilities for the failure time of the next unit based on such data.
19

A key motivation for this work, as alternative to our recent NPI-methods for age replacement based on the renewal criterion [6, 7], is the natural use of the one-cycle criterion for
fully adaptive methods, with the idea that the replacement time for unit n+2 will also adapt
to the information that we would get on unit n + 1 within the process. We have not studied
this adaptation for further future units in this paper, but we did consider it explicitly in
[7] for the renewal criterion. Clearly, for smaller n the extra information from unit n + 1
has relatively more influence on the optimal replacement time for unit n + 2 than for larger
values of n. As far as we are aware, no detailed comparisons of different optimality criteria
for age replacement have been published. In practice, it may not be precisely clear which
criterion is most suitable, there may well be other criteria that fit better with some particular
applications. In addition to direct comparisons of optimal replacement times according to
different criteria, as presented in Sections 2 and 5, it is also of interest to study robustness
of replacement strategies with regard to different optimality criteria, both within NPI and
more generally.
Of course, in the literature many variations to age replacement have been studied, and
many of these can also be studied from NPI perspective. One such variation is so-called
opportunity-based age replacement [14], where it is acknowledged that one may not be able
to plan preventive replacement at any moment, but can only perform such replacement at
randomly occurring opportunities. In [15] we studied this from NPI perspective, restricting
to the renewal criterion, where again the methods seem to provide good adaptation to the
failure data.

20

Appendix
Proof of Lemma 1
a. For T = x(j) and j = 1, . . . , n + 1, we have
Cone (x(j) ) =

Zx(1)

c2
fX (x)dx +
x n+1

x(l+1)
j1 Z
X
c2
l=1 x
(l)

x(1)

= c2

j1

x(1)

= c2

2 x

dx +

j1
X
l=1

Zr

fXn+1 (x)dx

x(j)

x(l+1)

X
1 2 x
xe dx +
x
l=1

c1
fXn+1 (x)dx +
x
T

c1
c2
1
1
dx +
x (n + 1) (x(l+1) x(l) )
T

x(l)

n+1j
n+1

x(l+1)

c2
(n + 1)(x(l+1) x(l) )

c1
1
dx +
x
x(j)

x(l)

n+1j
n+1

j1


)
X ln(x(l+1) ) ln(x(l) )
n
+
1

j
1
c2
. (35)
+ c1
= c2 (1 ex(1) ) +
n+1
x
x
(l+1) x(l)
(j)
l=1
b. For T (x(j) , x(j+1) ) and j = 1, . . . , n, substituting (20) into
Cone (T ) =

Zx(1)
0

j1

X
c2
fXn+1 (x)dx +
x
l=1

x(l+1)

c2
fX (x)dx +
x n+1

Zr

c1
fX (x)dx +
T n+1

c1
fX (x)dx
T n+1

ZT

c2
fX (x)dx
x n+1

x(j)

x(l)

x(j+1)

(36)

x(j+1)

yields the expression given in Lemma 1b.

c. For T (0, x(1) ), substituting (20) into
Cone (T ) =

ZT
0

c2
fX (x)dx +
x n+1

Zx(1)
T

yields the expression given in Lemma 1c.

21

c1
fX (x)dx +
T n+1

Zr

x(1)

c1
fX (x)dx
T n+1

(37)

Proof of Lemma 2

a. For T (x(j) , x(j+1) ), j = 1, . . . , n, we have





x(j+1) T
c1
1
c1
dCone (T )
c2
+nj
=

dT
n + 1 T (x(j+1) x(j) ) T 2 x(j+1) x(j)
T (x(j+1) x(j) )


c1 x(j+1) c1 (n j)(x(j+1) x(j) )
1

c2
(38)
=
(n + 1)T (x(j+1) x(j) )
T
T
Hence, we have that

dCone (T )
= 0 where
dT
T =j
j =

while

c1
(x(j+1) + (n j)(x(j+1) x(j) ))
c2

(39)

dCone (T )
dCone (T )
< 0 for T (x(j) , j ) and
> 0 for T (j , x(j+1) ) as long as
dT
dT
c1
(40)
x(j) < (x(j+1) + (n j)(x(j+1) x(j) )) < x(j+1) ,
c2

that is, as long as the ratio c2 /c1 satisfies the condition given in Lemma 2a.
b. For T (0, x(1) ), we have
dCone (T )
1
c1
c1
c1
= c2 2 eT c1 2 eT 2 eT eT + 2 ex(1) (x(1) + )
dT
T
T
T

c1 n
2
T 
n+1
c1 
c1
T
(41)
= e
(c2 c1 )2 2
T
T

where we have used the fact that has been chosen such that it satisfies (17). Hence,
dCone (T )
we have that
= 0 where
dT
T =0
p
c1 + c1 (4c2 3c1 )
0 =
(42)
2(c2 c1 )
dCone (T )
dCone (T )
< 0 for T (0, 0 ) and
> 0 for T (0 , x(1) ) as long as
dT
dT
0 < 0 < x(1) .

while

Acknowledgement
This research is supported by the UK Engineering and Physical Research Sciences Council,
grant GR/R92530/01.
22

References
[1] Barlow, R.E. and Proschan, F. (1965). Mathematical Theory of Reliability. Wiley, New
York.
[2] Aven, T. and Jensen, U. (1999). Stochastic Models in Reliability. Springer, New York.
[3] Mazzuchi, T.A. and Soyer, R. (1996). A Bayesian perspective on some replacement
strategies. Reliability Engineering and System Safety, 51, 295-303.
[4] Sheu, S.H., Yeh, R.H., Lin, Y.B. and Juan, M.G. (1999). A Bayesian perspective on age
replacement with minimal repair. Reliability Engineering and System Safety, 65, 55-64.
[5] Sheu, S.H., Yeh, R.H., Lin, Y.B. and Juan, M.G. (2001). A Bayesian approach to an
adaptive preventive maintenance model. Reliability Engineering and System Safety, 71,
33-44.
[6] Coolen-Schrijner, P. and Coolen, F.P.A. (2004a). Non-parametric predictive inference
for age replacement with a renewal argument. Quality and Reliablity Engineering International, 20, 203-215.
[7] Coolen-Schrijner, P. and Coolen, F.P.A. (2004b). Adaptive age replacement strategies
based on nonparametric predictive inference. Journal of the Operational Research Society, to appear.
[8] Coolen, F.P.A., Coolen-Schrijner, P. and Yan, K.J. (2002). Nonparametric predictive
inference in reliability. Reliability Engineering and System Safety, 78, 185-193.
[9] Hill, B.M. (1968). Posterior distribution of percentiles: Bayes theorem for sampling
from a population. Journal of the American Statistical Association, 63, 677-691.
[10] Augustin, T. and Coolen, F.P.A. (2004). Nonparametric predictive inference and interval
probability. Journal of Statistical Planning and Inference, 124, 251-272.
[11] Luenberger, D.G. (1973). Introduction to Linear and Nonlinear Programming. AddisonWesley, Menlo Park.
[12] The R Project for Statistical Computing: www.r-project.org.
[13] Coolen, F.P.A. and Yan, K.J. (2004). Nonparametric predictive inference with rightcensored data. Journal of Statistical Planning and Inference, 126, 25-54.
23

[14] Dekker, R. and Dijkstra, M.C. (1992). Opportunity-based age replacement: Exponentially distributed times between opportunities. Naval Research Logistics, 39, 175-190.
[15] Coolen-Schrijner, P., Coolen, F.P.A. and Shaw, S.C. (2004). Nonparametric adaptive
opportunity-based age replacement strategies. In submission.

24